You can deploy this model and start earning money today!

amkyawdev/
Myanmar-Ghost-Instruct-LoRA

Name: Myanmar-Ghost-Instruct-LoRA
Availability: InStock

Qwen/Qwen2.5-Coder-1.5B-Instruct

apache-2.0

Views

No views yet

🇲🇲 Myanmar-Ghost-Instruct-LoRA

Myanmar Language Instruction-Tuned LLM based on Qwen2.5-Coder-1.5B-Instruct

A lightweight LoRA adapter for Myanmar language text generation and instruction following

📌 Model Overview

Myanmar-Ghost-Instruct-LoRA is a LoRA (Low-Rank Adaptation) adapter trained on Qwen/Qwen2.5-Coder-1.5B-Instruct to enhance Myanmar (Burmese) language understanding and generation capabilities.

Key Features

🐍 Lightweight: Only ~74MB (LoRA adapter)
🇲🇲 Myanmar-First: Optimized for Burmese text generation
💻 Code Capable: Base model retains code generation abilities
⚡ Fast Inference: Low-rank adaptation for efficient deployment
🔧 Easy Integration: Compatible with PEFT and Transformers libraries

Model Tree

Qwen/Qwen2.5-1.5B
└── Qwen/Qwen2.5-Coder-1.5B
    └── Qwen/Qwen2.5-Coder-1.5B-Instruct
        └── amkyawdev/Myanmar-Ghost-Instruct-LoRA ✅ (this model)

🚀 Quick Start

Using PEFT (Recommended)

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load base model and tokenizer
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-Coder-1.5B-Instruct",
    device_map="auto",
    torch_dtype=torch.float16,
    trust_remote_code=True
)

tokenizer = AutoTokenizer.from_pretrained(
    "Qwen/Qwen2.5-Coder-1.5B-Instruct",
    trust_remote_code=True
)

# Load LoRA adapter
model = PeftModel.from_pretrained(
    base_model, 
    "amkyawdev/Myanmar-Ghost-Instruct-LoRA"
)

# Generate text
messages = [
    {"role": "user", "content": "မြန်မာစာတစ်ပိုဒ် ရေးပါ။"}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.9
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Using Transformers Pipeline

python

from transformers import pipeline

pipe = pipeline(
    "text-generation",
    model="amkyawdev/Myanmar-Ghost-Instruct-LoRA",
    model_kwargs={"device_map": "auto", "torch_dtype": "float16"}
)

messages = [
    {"role": "user", "content": "မြန်မာစာတစ်ပိုဒ် ရေးပါ။"}
]
output = pipe(messages, max_new_tokens=512, temperature=0.7)
print(output[0]["generated_text"])

Using vLLM

bash

# Install vLLM
pip install vllm

# Start server
vllm serve "amkyawdev/Myanmar-Ghost-Instruct-LoRA" --dtype float16

# API call
curl -X POST "http://localhost:8000/v1/chat/completions" \
  -H "Content-Type: application/json" \
  --data '{
    "model": "amkyawdev/Myanmar-Ghost-Instruct-LoRA",
    "messages": [{"role": "user", "content": "မြန်မာစာတစ်ပိုဒ် ရေးပါ။"}]
  }'

📊 Technical Specifications

LoRA Configuration

Parameter	Value
PEFT Type	LORA
Rank (r)	16
Alpha	32
Dropout	0.05
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Bias	none
Task Type	CAUSAL_LM

Tokenizer

Property	Value
Tokenizer Class	Qwen2Tokenizer
Model Max Length	32,768 tokens
Special Tokens	`<
Padding Token	`<

Adapter File Size

adapter_model.safetensors: ~74 MB
Total model size (with base): ~3-4 GB

🏋️ Training Details

Training Hyperparameters

Parameter	Value
Total Steps	200
Save Steps	100
Batch Size	1
Max Steps per Epoch	12,500 (estimated)
Training Epochs	1
Max Learning Rate	2e-4 (warmup)
Final Learning Rate	~5.2e-8
Training Framework	PEFT 0.19.1
Base Model	Qwen2.5-Coder-1.5B-Instruct

Training Progress

Step	Loss	Learning Rate	Grad Norm
1	12.01	0.0	6.93
50	~2.5	~1e-4	~3.0
100	~1.5	~5e-5	~2.5
200 (final)	1.91	5.2e-8	2.66

Available Checkpoints

checkpoint-100/ - Model at step 100
checkpoint-200/ - Final model at step 200

📚 Training Data

This model was trained on Myanmar language instruction datasets including:

Myanmar V3 Clean Dataset (amkyawdev/myanmar-v3-clean)
- ~878K samples
- Cleaned and quality-filtered Myanmar text
AMK Coder V3 Dataset V2 (amkyawdev/amk-coder-v3-dataset-v2)
- ~1.01M samples
- Code and natural language instruction pairs

🎯 Intended Uses

Direct Use Cases

✅ Myanmar language text generation
✅ Burmese language conversation
✅ Translation assistance (Myanmar ↔ other languages)
✅ Text summarization in Burmese
✅ Code generation assistance (preserved from base model)

Downstream Use Cases

🔧 Fine-tuning for specific Myanmar NLP tasks
🔧 Domain-specific applications (healthcare, legal, education)
🔧 Chatbot development for Burmese speakers
🔧 Research on low-resource language LLMs

Out-of-Scope Uses

⚠️ Medical or legal advice without human verification
⚠️ High-stakes decision-making systems
⚠️ Production systems without thorough evaluation
⚠️ Generating harmful or misleading content

⚠️ Bias, Risks, and Limitations

Technical Limitations

Model Size: 1.5B parameters may limit performance on complex tasks
Training Steps: Limited training (200 steps) may affect instruction-following quality
Token Limit: 32,768 context window
Resource Requirements: GPU recommended for inference

Sociotechnical Considerations

Language Coverage: Optimized primarily for Burmese; may vary for regional dialects
Cultural Bias: Training data may reflect specific cultural perspectives
Safety: As with any LLM, outputs should be verified before critical use

Recommendations

Evaluate on your specific use case before production deployment
Implement appropriate content filtering
Provide human oversight for sensitive applications
Consider fine-tuning for domain-specific tasks

📈 Evaluation

Evaluation Status

⚠️ Formal benchmark evaluation pending. The model has not been systematically evaluated on standard NLP benchmarks yet.

Recommended Evaluation Tasks

If you evaluate this model, consider the following benchmarks:

Myanmar NLP Tasks
- Myanmar text classification
- Sentiment analysis (Burmese)
- Named entity recognition
General Language Tasks
- MMLU (Multilingual Massive Multitask)
- Hellaswag
- TruthfulQA
Code Generation (inherited from base model)
- HumanEval
- MBPP

User Feedback

We welcome community feedback! Please share your evaluation results and use cases in the Discussions tab.

🔧 Merge and Deploy

Merge LoRA with Base Model

python

from peft import PeftModel
from transformers import AutoModelForCausalLM
import torch

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-Coder-1.5B-Instruct",
    device_map="cpu",
    torch_dtype=torch.float32,
)
model = PeftModel.from_pretrained(base_model, "amkyawdev/Myanmar-Ghost-Instruct-LoRA")

# Merge adapter weights
merged_model = model.merge_and_unload()
merged_model.save_pretrained("merged-model")

Quantization for Deployment

python

# 4-bit quantization with GGUF
from transformers import AutoModelForCausalLM, BitsAndBytesConfig

quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16
)

model = AutoModelForCausalLM.from_pretrained(
    "amkyawdev/Myanmar-Ghost-Instruct-LoRA",
    quantization_config=quantization_config,
    device_map="auto"
)

🌐 Related Models

Explore more models from the author:

Model	Description
Myanmar-Ghost-Instruct-GGUF	GGUF format for local inference
myanmar-ai-v3	Full model version
qwen2.5-myanmar-ai-adapter	Alternative adapter
amk-coder-v2	Coding-focused model

📞 Contact & Support

Framework Versions

PEFT: 0.19.1
Transformers: Compatible with latest version
PyTorch: Recommended 2.0+

📄 License

This adapter is released under the Apache 2.0 License.

The base model Qwen/Qwen2.5-Coder-1.5B-Instruct is licensed by Alibaba Cloud and subject to its terms.

Made with ❤️ for the Myanmar AI community

This model card was created to improve transparency and reproducibility.